v0.41

What's new (v0.41)

No more re-downloading the model in a loop. On a rig where AI inference fails (for example a
GPU that produces NaN during inference), the miner used to delete the cached model as "corrupt" and
re-download the full multi-GB file again on every restart — a perpetual loop that never let the rig
settle. That's fixed.

🔁 One-shot self-heal — model is re-fetched at most once, ever

The self-heal that deletes and re-downloads a model on a corrupt-model error is now capped by an
on-disk marker, not just in memory. So even if the miner is restarted (e.g. by HiveOS) between
attempts, the model is re-downloaded at most once. If it still fails after that, the weights are
kept in place and PoW mining keeps running at full speed — only AI inference for that one model
stays disabled, instead of burning bandwidth re-downloading 5 GB on every restart.

If a later attempt does produce a valid inference, the marker is cleared automatically, so a
genuinely fixed file can still self-heal in the future.

If AI inference misbehaves on a particular GPU, you can also run with --cpu-inference to keep
the GPU free for hashing and run inference on the CPU.

Carried forward

• v0.40 — warm reconnects (GPUs keep hashing across dropped pool connections).
• v0.39 — chunked ~512 MB model downloads. v0.38 — --high/--very-high on the pool.
• v0.37 resumable downloads · v0.36 OPoI results reach the pool · v0.35 model self-heal.

---

What's new (v0.40)

Mining survives dropped connections. On an unstable or long-haul link (where the pool
connection gets reset every so often), the miner no longer tears down and re-initialises your GPUs
on every reconnect — they stay warm and keep hashing while the stratum socket reconnects underneath.

🔌 Warm reconnects — no more GPU re-init on every drop

Previously a single "Connection reset by peer" closed every GPU thread, re-created the CUDA context,
and started the hashrate from zero again — several seconds of idle plus a ramp on each reconnect.
On a flaky link that adds up fast (and hits multi-GPU rigs hardest).

Now the GPU/CPU workers are spawned once and kept alive for the whole session. Their solutions flow
through a persistent relay that simply re-points at each freshly-reconnected pool connection. A
dropped connection now logs reconnecting (GPU workers kept warm) and mining continues without a
teardown or a hashrate ramp.

Verified on a live rig: forcing a mid-session disconnect left the GPU threads running (no
re-spawn) and accepted shares kept climbing straight through the reconnect.

Carried forward

• v0.39 — chunked ~512 MB model downloads (no Cloudflare mid-stream cut-offs).
• v0.38 — --high / --very-high tiers for pool mining. v0.37 — resumable downloads.
• v0.36 OPoI results reach the pool · v0.35 model fix + self-heal · v0.32 RTX 50-series.

---

What's new (v0.39)

Release artifacts

• mc-miner-v0.39-linux-ubuntu22-x86_64.tar.gz
• mc-miner-v0.39-linux-ubuntu24-x86_64.tar.gz
• mc-miner-v0.39-linux-hiveos.tar.gz
• mc-miner-v0.39-windows-x86_64.zip

Faster, reliable model downloads. Models now download in fixed ~512 MB chunks, so the first-run
download finishes cleanly even on fast links — no more mid-stream cut-offs.

⬇️ Chunked downloads

Large model files served through Cloudflare were getting cut off near ~1.1 GB on a single stream
(independent of link speed). The miner now fetches each model in bounded ~512 MB ranges that always
complete, appending them into place. v0.37's resume stays as a safety net for transient drops.

Result: a fresh rig pulls the full model in one smooth pass — no failed-transfer retries in the log.

Carried forward

• v0.38 — --high / --very-high model tiers work for pool mining (bigger models, bigger
  AI-inference rewards on high-VRAM rigs).
• v0.37 — resumable model downloads. v0.36 — OPoI results reach the pool.
• v0.35 model fix + self-heal · v0.34 hashrate report · v0.33 xbinor.io gateway ·
  v0.32 RTX 50-series fix.

---

What's new (v0.38)

Release artifacts

• mc-miner-v0.38-linux-ubuntu22-x86_64.tar.gz
• mc-miner-v0.38-linux-ubuntu24-x86_64.tar.gz
• mc-miner-v0.38-linux-hiveos.tar.gz
• mc-miner-v0.38-windows-x86_64.zip

--high / --very-high model tiers now work for pool mining too. Recommended if you have a
high-VRAM rig and want to answer the larger (higher-reward) AI inference requests.

🧠 Pool mining now honors the model-tier flags

Previously, mining to the pool always loaded only DeepSeek-R1-8B and ignored --light / --high
/ --very-high — so a 24 GB+ or 48 GB+ rig could not serve the 32B / 70B models or answer the
larger on-chain AiRequests (which carry bigger inference rewards).

Now the tier flags take effect for pool and solo alike:

--light TinyLlama-1.1B (6 GB+ VRAM)
(default) DeepSeek-R1-8B (12 GB+, pool's OPoI model)
--high + DeepSeek-R1-32B (24 GB+)
--very-high + Llama-3.3-70B (48 GB+ single GPU)

The pool default (no flag) stays DeepSeek-R1-8B. Bigger models download on first use.

Carried forward

• v0.37 — resumable model downloads (survive Cloudflare cut-offs on slow links).
• v0.36 — OPoI inference results reach the pool. v0.35 — model fix + self-heal.
• v0.34 hashrate report · v0.33 xbinor.io gateway · v0.32 RTX 50-series fix.

v0.40

What's new (v0.40)

Mining survives dropped connections. On an unstable or long-haul link (where the pool
connection gets reset every so often), the miner no longer tears down and re-initialises your GPUs
on every reconnect — they stay warm and keep hashing while the stratum socket reconnects underneath.

🔌 Warm reconnects — no more GPU re-init on every drop

Previously a single "Connection reset by peer" closed every GPU thread, re-created the CUDA context,
and started the hashrate from zero again — several seconds of idle plus a ramp on each reconnect.
On a flaky link that adds up fast (and hits multi-GPU rigs hardest).

Now the GPU/CPU workers are spawned once and kept alive for the whole session. Their solutions flow
through a persistent relay that simply re-points at each freshly-reconnected pool connection. A
dropped connection now logs reconnecting (GPU workers kept warm) and mining continues without a
teardown or a hashrate ramp.

Verified on a live rig: forcing a mid-session disconnect left the GPU threads running (no
re-spawn) and accepted shares kept climbing straight through the reconnect.

Carried forward

• v0.39 — chunked ~512 MB model downloads (no Cloudflare mid-stream cut-offs).
• v0.38 — --high / --very-high tiers for pool mining. v0.37 — resumable downloads.
• v0.36 OPoI results reach the pool · v0.35 model fix + self-heal · v0.32 RTX 50-series.

---

What's new (v0.39)

Release artifacts

• mc-miner-v0.39-linux-ubuntu22-x86_64.tar.gz
• mc-miner-v0.39-linux-ubuntu24-x86_64.tar.gz
• mc-miner-v0.39-linux-hiveos.tar.gz
• mc-miner-v0.39-windows-x86_64.zip

Faster, reliable model downloads. Models now download in fixed ~512 MB chunks, so the first-run
download finishes cleanly even on fast links — no more mid-stream cut-offs.

⬇️ Chunked downloads

Large model files served through Cloudflare were getting cut off near ~1.1 GB on a single stream
(independent of link speed). The miner now fetches each model in bounded ~512 MB ranges that always
complete, appending them into place. v0.37's resume stays as a safety net for transient drops.

Result: a fresh rig pulls the full model in one smooth pass — no failed-transfer retries in the log.

Carried forward

• v0.38 — --high / --very-high model tiers work for pool mining (bigger models, bigger
  AI-inference rewards on high-VRAM rigs).
• v0.37 — resumable model downloads. v0.36 — OPoI results reach the pool.
• v0.35 model fix + self-heal · v0.34 hashrate report · v0.33 xbinor.io gateway ·
  v0.32 RTX 50-series fix.

---

What's new (v0.38)

Release artifacts

• mc-miner-v0.38-linux-ubuntu22-x86_64.tar.gz
• mc-miner-v0.38-linux-ubuntu24-x86_64.tar.gz
• mc-miner-v0.38-linux-hiveos.tar.gz
• mc-miner-v0.38-windows-x86_64.zip

--high / --very-high model tiers now work for pool mining too. Recommended if you have a
high-VRAM rig and want to answer the larger (higher-reward) AI inference requests.

🧠 Pool mining now honors the model-tier flags

Previously, mining to the pool always loaded only DeepSeek-R1-8B and ignored --light / --high
/ --very-high — so a 24 GB+ or 48 GB+ rig could not serve the 32B / 70B models or answer the
larger on-chain AiRequests (which carry bigger inference rewards).

Now the tier flags take effect for pool and solo alike:

--light TinyLlama-1.1B (6 GB+ VRAM)
(default) DeepSeek-R1-8B (12 GB+, pool's OPoI model)
--high + DeepSeek-R1-32B (24 GB+)
--very-high + Llama-3.3-70B (48 GB+ single GPU)

The pool default (no flag) stays DeepSeek-R1-8B. Bigger models download on first use.

Carried forward

• v0.37 — resumable model downloads (survive Cloudflare cut-offs on slow links).
• v0.36 — OPoI inference results reach the pool. v0.35 — model fix + self-heal.
• v0.34 hashrate report · v0.33 xbinor.io gateway · v0.32 RTX 50-series fix.

v0.39

What's new (v0.39)

Release artifacts

• mc-miner-v0.39-linux-ubuntu22-x86_64.tar.gz
• mc-miner-v0.39-linux-ubuntu24-x86_64.tar.gz
• mc-miner-v0.39-linux-hiveos.tar.gz
• mc-miner-v0.39-windows-x86_64.zip

Faster, reliable model downloads. Models now download in fixed ~512 MB chunks, so the first-run
download finishes cleanly even on fast links — no more mid-stream cut-offs.

⬇️ Chunked downloads

Large model files served through Cloudflare were getting cut off near ~1.1 GB on a single stream
(independent of link speed). The miner now fetches each model in bounded ~512 MB ranges that always
complete, appending them into place. v0.37's resume stays as a safety net for transient drops.

Result: a fresh rig pulls the full model in one smooth pass — no failed-transfer retries in the log.

Carried forward

• v0.38 — --high / --very-high model tiers work for pool mining (bigger models, bigger
  AI-inference rewards on high-VRAM rigs).
• v0.37 — resumable model downloads. v0.36 — OPoI results reach the pool.
• v0.35 model fix + self-heal · v0.34 hashrate report · v0.33 xbinor.io gateway ·
  v0.32 RTX 50-series fix.

---

What's new (v0.38)

Release artifacts

• mc-miner-v0.38-linux-ubuntu22-x86_64.tar.gz
• mc-miner-v0.38-linux-ubuntu24-x86_64.tar.gz
• mc-miner-v0.38-linux-hiveos.tar.gz
• mc-miner-v0.38-windows-x86_64.zip

--high / --very-high model tiers now work for pool mining too. Recommended if you have a
high-VRAM rig and want to answer the larger (higher-reward) AI inference requests.

🧠 Pool mining now honors the model-tier flags

Previously, mining to the pool always loaded only DeepSeek-R1-8B and ignored --light / --high
/ --very-high — so a 24 GB+ or 48 GB+ rig could not serve the 32B / 70B models or answer the
larger on-chain AiRequests (which carry bigger inference rewards).

Now the tier flags take effect for pool and solo alike:

--light TinyLlama-1.1B (6 GB+ VRAM)
(default) DeepSeek-R1-8B (12 GB+, pool's OPoI model)
--high + DeepSeek-R1-32B (24 GB+)
--very-high + Llama-3.3-70B (48 GB+ single GPU)

The pool default (no flag) stays DeepSeek-R1-8B. Bigger models download on first use.

Carried forward

• v0.37 — resumable model downloads (survive Cloudflare cut-offs on slow links).
• v0.36 — OPoI inference results reach the pool. v0.35 — model fix + self-heal.
• v0.34 hashrate report · v0.33 xbinor.io gateway · v0.32 RTX 50-series fix.

v0.38

What's new (v0.38)

Release artifacts

• mc-miner-v0.38-linux-ubuntu22-x86_64.tar.gz
• mc-miner-v0.38-linux-ubuntu24-x86_64.tar.gz
• mc-miner-v0.38-linux-hiveos.tar.gz
• mc-miner-v0.38-windows-x86_64.zip

--high / --very-high model tiers now work for pool mining too. Recommended if you have a
high-VRAM rig and want to answer the larger (higher-reward) AI inference requests.

🧠 Pool mining now honors the model-tier flags

Previously, mining to the pool always loaded only DeepSeek-R1-8B and ignored --light / --high
/ --very-high — so a 24 GB+ or 48 GB+ rig could not serve the 32B / 70B models or answer the
larger on-chain AiRequests (which carry bigger inference rewards).

Now the tier flags take effect for pool and solo alike:

--light TinyLlama-1.1B (6 GB+ VRAM)
(default) DeepSeek-R1-8B (12 GB+, pool's OPoI model)
--high + DeepSeek-R1-32B (24 GB+)
--very-high + Llama-3.3-70B (48 GB+ single GPU)

The pool default (no flag) stays DeepSeek-R1-8B. Bigger models download on first use.

Carried forward

• v0.37 — resumable model downloads (survive Cloudflare cut-offs on slow links).
• v0.36 — OPoI inference results reach the pool. v0.35 — model fix + self-heal.
• v0.34 hashrate report · v0.33 xbinor.io gateway · v0.32 RTX 50-series fix.

v0.37

What's new (v0.37)

Reliable model downloads on slow / unstable connections. Recommended especially if your first-run
model download keeps failing.

⬇️ Fix: resumable model downloads

Large model files are served through Cloudflare, which can cut off a long download mid-stream on
slower links:

[keryx-miner] Downloading https://xbinor.io/ipfs/... 1120.3/4920.7 MB (22%)
IPFS: gateway ... failed: response body closed before all bytes were read

The miner used to restart from 0 on every cut-off, so a slow connection could never finish. Now it
streams into a .part file and resumes from the bytes already on disk (Range request) on each
reconnect, retrying while progress is made. A dropped 4.9 GB download now continues instead of
starting over — it completes even on a slow link.

Carried forward

• v0.36 — OPoI inference results reach the pool (delivered on the next share, matched by
  challenge id).
• v0.35 — DeepSeek-R1-8B model fix + self-healing model downloads.
• v0.34 — hashrate reported to the pool. v0.33 — xbinor.io gateway. v0.32 — RTX 50-series fix.

v0.36

What's new (v0.36)

Completes the OPoI AI-inference round-trip: inference results now actually reach the pool.
Recommended for everyone; no config changes needed.

🔗 Fix: inference results now reach the pool

A miner's AI inference result is computed for a specific challenge, but SLM inference takes a few
seconds while mining jobs rotate sub-second — so by the time the result was ready, the miner only
tried to attach it to a share of the (long-gone) original job, and it was never sent. The result
was stranded on every challenge.

Now the miner delivers the latest ready result on the next share regardless of which job it is
mining, carrying the original challenge id so the pool matches the right request. End-to-end the
OPoI inference round-trip works: the pool can publish the on-chain AiResponse and credit the
inference reward to the solving rig.

Carried forward

• v0.35 — DeepSeek-R1-8B inference fix (correct model) + self-healing model downloads.
• v0.34 — hashrate reported to the pool (dashboard matches the console).
• v0.33 — faster model downloads via the xbinor.io gateway (Keryx fallback).
• v0.32 — RTX 50-series (Blackwell / sm_120) fix.
• v0.31 / v0.30 — high-GPU-count crash fix; pool-proxy OPoI; no dev fee; DAA-gated SALT parity.

v0.35

What's new (v0.35)

Fixes broken OPoI inference on DeepSeek-R1-8B and makes the miner self-heal bad model files.
Recommended for everyone; no config changes needed.

🧠 Fix: DeepSeek-R1-8B inference (and self-healing model downloads)

OPoI inference on deepseek-r1-8b was failing on every challenge with:

SlmEngine 'deepseek-r1-8b' generate error: sample: A weight is negative, too large or not a valid number

The cause was a defectively quantized model file — its weights produced NaN/Inf during
inference, so no valid result was ever produced (mining was unaffected; only the AI inference
result was broken). Fixed on two fronts:

• the model file has been replaced with a correctly quantized one (same size, downloads from
  the same URL), so fresh installs just work; and
• the miner now self-heals: if a loaded model produces NaN/Inf logits, or fails to load with
  an unsupported quantization, it deletes the cached file and re-downloads it once — so rigs
  that already cached the bad file recover automatically on this update, no manual cleanup needed.

The miner also no longer submits an error string as an "inference result", and logs OPoI
challenges honestly (whether a result was attached or not).

Already running and stuck on the old bad model? Just update to v0.35 and restart — it
re-downloads the fixed file on its own. (Or delete models/DeepSeek-R1-8B/ and restart.)

Carried forward (from v0.34)

• Hashrate reported to the pool (mining.submit_hashrate) so the dashboard matches the console.

Carried forward (from v0.33)

• Faster model downloads via the xbinor.io IPFS gateway, with the Keryx gateway as a fallback.

Carried forward (from v0.32)

• RTX 50-series (Blackwell / sm_120) fix — broad-driver-compatible GPU kernels, a kernel
  fallback chain, and a driver-version gate that logs clearly instead of unknown error.

Carried forward (from v0.31 / v0.30)

• High-GPU-count crash fix; self-healing tokenizer download.
• Pool-proxy OPoI inference; no miner dev fee (100% to your address); DAA-gated SALT parity;
  on-demand model tiers (--high / --very-high / --cpu-inference).

v0.34

What's new (v0.34)

The pool dashboard now shows the same hashrate as the miner console. Recommended for everyone;
no config changes needed.

📊 Hashrate reported to the pool

The miner now reports its measured hashrate to the pool (mining.submit_hashrate) every 30
seconds, so the pool's per-worker hashrate matches what the console prints — instead of the
pool's share-based estimate, which lags and varies short-term. Each rig reports its own rate;
the pool sums them per address as before.

• No action needed — it just works once you update.

Carried forward (from v0.33)

• Faster model downloads via the xbinor.io IPFS gateway, with the Keryx gateway as an automatic
  fallback (--ipfs-gateway to override).

Carried forward (from v0.32)

• RTX 50-series (Blackwell / sm_120) fix — GPU kernels built for broad driver compatibility,
  a kernel fallback chain, and a driver-version gate that logs clearly instead of unknown error.

Carried forward (from v0.31)

• Crash fix for high-GPU-count rigs (share-pending lock now waits instead of panicking).
• Self-healing model download — a truncated/corrupt tokenizer is re-fetched automatically.

Carried forward (from v0.30)

• Pool-proxy OPoI inference — no local IPFS, inference reward credited to your rig.
• No miner dev fee (mines 100% to your address); bounded inference + reliable auto-reconnect.
• DAA-gated SALT v1 → v2 → v4 consensus parity; per-share OPoI proof (required by default);
  on-demand model tiers (--high / --very-high / --cpu-inference).

v0.33

What's new (v0.33)

Faster, more reliable model downloads. Recommended for everyone; no config changes needed.

🌐 Model downloads via a faster gateway (with automatic fallback)

First-run model downloads now use a faster primary IPFS gateway, with the public Keryx
gateway kept as an automatic fallback — if the primary is unreachable, the miner retries
the next gateway instead of failing. Each model file is fetched from the gateways in order
until one succeeds, so a single gateway hiccup no longer stops a rig from syncing.

• Advanced: override the primary gateway with --ipfs-gateway https://your-gateway (the Keryx
  gateway stays as a fallback). No effect on rigs that already have their models downloaded.

Carried forward (from v0.32)

• RTX 50-series (Blackwell / sm_120) fix — GPU kernels built for broad driver compatibility,
  a kernel fallback chain, and a driver-version gate that logs clearly instead of
  unknown error. Blackwell rigs on CUDA 12.8+ drivers mine out of the box.

Carried forward (from v0.31)

• Crash fix for high-GPU-count rigs (share-pending lock now waits instead of panicking).
• Self-healing model download — a truncated/corrupt tokenizer is re-fetched automatically.

Carried forward (from v0.30)

• Pool-proxy OPoI inference — no local IPFS, inference reward credited to your rig.
• No miner dev fee (mines 100% to your address); bounded inference + reliable auto-reconnect.
• DAA-gated SALT v1 → v2 → v4 consensus parity; per-share OPoI proof (required by default);
  on-demand model tiers (--high / --very-high / --cpu-inference).

v0.32

What's new (v0.32)

Fix for RTX 50-series (Blackwell) rigs that couldn't start mining at all. Strongly
recommended for anyone on an RTX 5070 / 5080 / 5090 or any other Blackwell GPU.

🛠️ Fix: RTX 50-series (sm_120) "failed to initialize"

On Blackwell cards the miner could disable the GPU at startup and drop into an endless
reconnect loop:

GPU #0 (NVIDIA GeForce RTX 5090) failed to initialize and was disabled:
CUDA PTX load failed for sm_120 ... "unknown error"
launching: 0 cpu miners
All workers are already dead

The bundled GPU code for sm_120/sm_100 was compiled with a CUDA toolkit newer than the
driver most rigs actually run, so the driver couldn't load it and the card was dropped. Fixed
on three levels:

• the sm_120 and sm_100 GPU kernels are now built for broad driver compatibility — they
  load on every Blackwell-capable driver, not just the very newest one;
• the miner now tries a chain of compatible kernels and uses the first one your driver can
  load, instead of giving up after the first failure; and
• if a kernel is too new for your driver, the log now says so clearly (e.g.
  needs CUDA driver >= 12.8 but this driver is 12.4) instead of a cryptic unknown error.

No config changes needed — just update and restart. If your RTX 50-series rig was stuck
reconnecting, this release gets it mining.

Carried forward (from v0.31)

• Crash fix for high-GPU-count rigs (share-pending lock now waits instead of panicking).
• Self-healing model download — a truncated/corrupt tokenizer is re-fetched automatically.

Carried forward (from v0.30)

• Pool-proxy OPoI inference — no local IPFS, inference reward credited to your rig.
• No miner dev fee (mines 100% to your address); bounded inference + reliable auto-reconnect.
• DAA-gated SALT v1 → v2 → v4 consensus parity; per-share OPoI proof (required by default);
  on-demand model tiers (--high / --very-high / --cpu-inference).